Multi-Scale Speaker Vectors for Zero-Shot Speech Synthesis
Tristin Cory, Razib Iqbal · 2022 IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC) · 2022
Recent advances in deep learning have allowed the task of synthesizing speech from text, otherwise known as Text-to-Speech (TTS), to be addressed with more powerful and effective techniques than previously possible. In this paper, we propose a novel speaker encoder model utilizing some of those techniques, such as self-attention and transformers. The proposed model results in a speaker embedding vector called Multi-Scale Speaker (MSS) vectors. MSS vectors aim to make improvements over the current state-of-the-art speaker embeddings for more natural and similar-sounding synthesized speech for unseen speakers in a zero-shot speech synthesis model. The proposed architecture relies on encoder layers which are coupled with a multi-scale approach for learning both local and global mel-spectra features of a reference speaker. We compared our proposed model against a modified generalized end-to-end (GE2E) speaker encoder using mel-cepstrum distortion and cosine similarity measures as well as mean opinion scores. Evaluation results confirm our encoder's advantages over a state-of-the-art speaker encoder model.